Microsoft Road Detections

Bing Maps found these roads by machine. Its imagery, which includes pictures from Maxar and Airbus, went through a neural network, and the roads it found are free to download.¹ Microsoft gives its reason for releasing them as support for the OpenStreetMap community.¹

The work happens in two steps.¹ A convolutional network, a kind of image model, marks which pixels in a 1 metre image are road.¹ A second stage of ordinary algorithms thins those pixels to centre lines, joins the gaps, builds a connected network and stitches tiles together.^1,4 The repository reports 85 per cent precision and 83 per cent recall for the pixel step, and says the roads that passed further checks reached 95 per cent precision before going into Microsoft's own maps.¹

The data comes as one zip file per region of the world, from Northern America at about 13 million kilometres to Eastern Asia at about 150 thousand.¹ Each line of a file holds a three-letter country code and one road as GeoJSON, with its approximate width in metres.¹ A road is as old as the imagery it was found in, and because Bing imagery is stitched from many sources, Microsoft says exact dates are hard to know.¹

For a model of a place, these roads can fill gaps in OpenStreetMap for access, evacuation or service planning. They are detections rather than a surveyed map, the only attribute is width, and a share of them will be wrong.¹

The road data is under the Open Database License (ODbL).^1,2 In October 2025 the team added the trained model, written for an old version of Keras, and the C# code that builds road lines from its output.^3,4,5 The repository's licence file speaks only of the data.²